run-logs / README.md
5xuekun's picture
S1B: index and card
04b5d1a verified
|
Raw History Blame Contribute Delete
4.32 kB
---
pretty_name: SHAPE run logs
license: cc-by-4.0
viewer: false
---
# SHAPE run logs
Run records behind the main and supplemental figures of the SHAPE manuscript: agent traces and LLM reasoning,
tool calls and outputs, scores, and checkpoints. Each top-level folder is named after the figure panels it supports.
Only runs that appear in a figure are included.
**Status (2026-11-14): complete.** All main and supplemental figure panels are covered.
Archive members of runs that no figure shows were removed.
## Folders
| Folder | Panels | Content |
| --- | --- | --- |
| `figure1e_reasoning_effort` | 1E | Batch correction (JUMP Target2, scenario 2) at low, medium and high reasoning effort, gpt5.6luna and sonnet 5.5 |
| `figure1g-i_crop` | 1D, 1G, 1I | Cropping on NeurIPS22-CellSeg (MaskCrop, ReAct, SHAPE; 1,000 images) and MedSegBench (334 images) |
| `figure1h_seg_neurips22` | 1D, 1H, S1A, S1C, S1D, S1E | Segmentation on NeurIPS22-CellSeg: StarDist, Cellpose3, microSAM, Cellpose-SAM, ReAct, SHAPE (gpt5.6luna); difficulty lists |
| `figure1j_ssl_rxrx1` | 1D, 1J, S1D, S1E | RxRx1 representation learning: DINOv3 fine-tune, ReAct and SHAPE (two LLM backends) for HepG2 and RPE; SHAPE HUVEC runs used in 1D; frozen-model records (DINOv3, DINO4Cells-B, DCM DINOv2-B, OpenPhenom) and the study summary |
| `figure1k_batch_jump` | 1D, 1F, 1K, S1D, S1E | Batch correction on JUMP Target2 (scenarios 2 and 4): Control preparation, ReAct and SHAPE (gpt5.6luna, high effort), and five further LLM backends (scenario 2) |
| `figure1l-n_e2e_jump` | 1M, 1N, S1F-S1I | End-to-end JUMP Cell Painting runs: the 600-image condition benchmark (`probe3`, seven arms, per-condition scoring) and the A549 CRISPR executions (`seed01`, `seed02`; all arms, CODEX in seed01 only) |
| `figure2d-h_cm_staining` | 2D-2H, S2A-S2E | cTnT staining runs (Control, CODEX, SHAPE; 2,070 images): records, per-chunk AnnData, embeddings, traces, the delivered SSL checkpoints, and the morphology-group analysis behind 2G and 2H |
| `figure2i-k_cm_flowcytometry` | 2I-2K, S2F | Flow-linked cardiomyocyte runs (Control, CODEX, SHAPE) |
| `figure3b_co_segmentation` | 3B | Cardiac organoid segmentation, 1,000 images, eight methods, AP over IoU thresholds |
| `figure3d-e_co_timecourse` | 3D, 3E, S3A | Cardiac organoid FC time course, 184 images (Control, CODEX, SHAPE) |
| `figure3f-g_co_drugscreen` | 3F, 3G | Cardiac organoid drug screen (SHAPE): run records, traces, the delivered SSL checkpoint and AnnData |
| `figure4b-e_tissues_celltypes` | 4B-4E, S4A, S4B | Xenium lung, ovary, prostate and skin: embeddings of Control, SHAPE, CODEX and scPortrait, three seeds, cell-class probes; transcript-defined cell-class labels |
| `figure4f-k_tumor_niches` | 4F-4K, S4C-S4F | Tumor-cell clusters, Hallmark scores, EMT niches and CellChat results (tables and code; the Xenium and MSigDB inputs are public and listed in `raw_data_sources/`) |
| `figureS1b_max_rounds` | S1B | SHAPE runs at increasing maximum round budgets for batch correction (JUMP), cropping and segmentation (NeurIPS22-CellSeg) and representation learning (RxRx1; caps 0, 1, 2, 4 and 6, with the delivered SSL checkpoint of each cap that made one), one archive per task with its member list |
## Provenance
Files come from `SHAPE-Morph/benchmark-logs` (revision `7c35e92d`) and `SHAPE-Morph/result` (revision `126eba1e`).
The S1B runs were not on Hugging Face before: their archives were packed from the run folders on the cluster, and
each `*.members.tsv` gives the size and SHA-256 of every file in its archive.
This repository keeps traces, scores, embeddings and checkpoints, not images: archives that held standalone image
files (segmentation masks, mask arrays, crops, plots) were repacked without them, and every other member of those
archives is byte-identical to the source. Images embedded inside trace files are left as recorded. All other files
are byte-identical copies.
`FIGURE_INDEX.tsv` lists, for every file, the panels it supports, the source repository, revision and path, the size,
its SHA-256 (large files) or git blob id (small files), and, for a repacked archive, how many image files were
removed and a digest of the kept members (SHA-256 of the sorted `name<TAB>sha256` lines). READMEs inside the folders
are copied from the source unchanged.